Experiment reports: paired-world tests for delegated work

[dongkeren]

Purpose

This thread is for runnable or manually reproducible tests of the six paired-world questions in the core runtime-semantics challenge. You may use KFD, another framework, a vendor runtime, or an independent harness.

The useful experimental shape is a pair of histories that should lead to different decisions but may collapse to the same state visible to the tested decision function.

Question IDs

• Q1 — Work Version: current authorized work versus a superseded version.
• Q2 — Authority Revocation: valid authority versus revoked or expired authority.
• Q3 — Causal History: evidence from the accepted causal path versus evidence from a stale attempt.
• Q4 — Retry Identity: a legitimate retry versus a duplicate consequential execution.
• Q5 — Recovery Drift: a valid recovery versus recovery after the objective, policy, or world has materially changed.
• Q6 — Accepted Completion: an output that exists versus a result whose completion was authoritatively accepted.

You may test one pair or all six. Preserve the question ID in the report so results can be compared across runtimes.

Minimal report template

Question ID: Q1 / Q2 / Q3 / Q4 / Q5 / Q6
Runtime or framework:
Version or source commit:
World A:
World B:
State visible to the decision function:
Expected decision in World A:
Expected decision in World B:
Observed decisions:
Discriminator, if any:
Where it is authoritatively bound:
Where may-advance is enforced:
Survives executor replacement: yes / no / unknown
Human reconstruction required: yes / no / unknown
Commands and public artifacts:
Claim boundary:

A failed or inconclusive run is useful. Passing a self-authored reference implementation is implementation evidence, not proof that every runtime has the same semantics.

Post a reply here for a small result, or open a separate Show and tell discussion for a substantial experiment. Formal counterevidence can be promoted to the KFD evidence form.